Dynamic tracking of non-stationarity in human ECoG activity.

Modeling brain network dynamics is essential in understanding neural mechanisms and developing neurotechnologies such as closed-loop stimulation therapies for a wide range of neurological disorders. Brain network activity could have non-stationary and time-variant dynamics, especially when the subject's brain is monitored for a long period, e.g., using the electrocorticogram (ECoG). This non-stationarity makes the modeling of dynamics challenging. In our prior work, we developed a framework to identify time-invariant linear state-space models (SSMs) to describe both stationary spontaneous neural population dynamics and input-output (IO) neural dynamics in response to electrical stimulation. Here, we develop an adaptive identification algorithm that estimates time-variant SSMs to track possible non-stationarity in brain network dynamics. We apply the adaptive algorithm to track high-density human ECoG dynamics in three subjects over a long time-period. We find that the adaptive identification algorithm can estimate time-variant SSMs that significantly outperform time-invariant SSMs in all subjects. Our results demonstrate that non-stationary dynamics exist in high-dimensional human ECoG signals over long time-periods, and that the proposed adaptive SSM identification algorithm can successfully track these non-stationarities. These results have important implications for more accurate estimation of neural biomarkers for different brain states and for adaptive closed-loop stimulation therapy across a wide range of neurological disorders.